Multi-Cancer Detection Using a Computationally Efficient CNN with Transfer Learning

Research Approach

This study addresses the challenge of deploying deep learning models for cancer detection in resource-constrained medical environments. The authors propose a lightweight convolutional neural network (CNN) architecture designed to minimize computational complexity. To overcome the limitations of small, labeled medical datasets, the researchers employ a transfer learning strategy. The model is first trained on one cancer dataset (source task) and then fine-tuned on others (target tasks), requiring only 20 additional epochs to achieve high performance. This approach aims to reduce the risk of overfitting and lower the computational overhead associated with retraining models for new clinical applications.

Key Findings

The proposed model demonstrates robust performance across three distinct tumor datasets: brain MRI, lung CT, and kidney CT. The model achieved test accuracies of 90.85% for brain cancer, 98.64% for lung cancer, and 99.92% for kidney cancer using 5-fold stratified cross-validation. The authors report that the fine-tuning process is highly efficient, requiring approximately 0.014 seconds per image per epoch on an NVIDIA GeForce GTX 960. Comparative analysis indicates that this lightweight architecture outperforms several state-of-the-art, more complex models, such as VGG16, VGG19, and DenseNet121, in terms of both efficiency and classification accuracy.

In clinical settings, medical imaging models often require frequent retraining to adapt to new scanners, institutions, or patient populations. Many existing deep learning frameworks are computationally intensive, making them difficult to implement in environments lacking high-end GPU infrastructure. By demonstrating that a streamlined, low-complexity CNN can achieve high diagnostic accuracy through transfer learning, this research provides a practical pathway for integrating AI-assisted diagnostics into routine clinical workflows without the need for excessive computational resources.